Dimension Reduction for Dynamic Analysis of Gene Expression Data Using Variational Autoencoder

Single-cell RNA sequencing technology captures gene expression at the whole-genome scale and single-cell level in many cells. Thus, the cellular states during a physiological process would be captured in detail. These data provide new opportunities and challenges for systems biology. Previously, the models in systems biology dealt with the dynamics of specific signaling pathways. However, it is almost impossible to directly model the dynamics of the whole genome due to the ultrahigh dimensionality of the data. Here, to model its gene regulatory dynamics, the dimensionality of the whole genome was reduced significantly using the latent space of a variational autoencoder, a generative deep learning model. In the low-dimensional latent space, an ordinary differentiation equation (ODE) system has been established to analyze the dynamic characteristics of the whole genome system. To demonstrate the effectiveness of this method, we generated a low-dimensional representation of a systems biology model we previously developed. The variational autoencoder reduced the dimension from 13 to 2 in latent space. An ODE model was developed to analyze the dynamics of this two-dimensional system. The results indicated that the 2-dimensional model captured the essence of the original systems biology model.